There's a moment every designer knows intimately, when you realise the tools you've spent years perfecting for others are utterly failing you personally. For me, that moment came in May 2023, staring at few dozen browser tabs and a bookmark folder labelled "Read Later" containing 2,847 unread articles.
The irony wasn't lost on me. I'd built user experiences for millions of people, but couldn't solve my own content consumption problem. Browser bookmarks had become digital graveyards. Pocket felt sterile. Instapaper worked but never quite sparked joy. We researched several other bookmarking apps, and the disconnect between saving content and actually engaging with it felt like a fundamental design challenge waiting to be solved.
That frustration spawned Jeto, an AI-powered bookmarking platform that would consume every spare moment for the next ten months. Along with my co-founder Sasi and product designer Stebin at Happy Machines Company, we set out to bridge the gap between digital hoarding and meaningful consumption.
We built something genuinely innovative. We never shipped it. And the lessons we learned are worth sharing for anyone considering AI-powered consumer products in 2024.
The core insight: bridging save and consume
The problem Jeto aimed to solve shows up in a simple behaviour observation: the average person bookmarks content 12 times more frequently than they revisit it. We're great digital hoarders but terrible digital consumers.
This isn't a willpower issue, it's a design opportunity. Traditional bookmarking tools optimise for capture but ignore the consumption journey entirely. If only Chrome bookmarks didn't feel like a dead end for links.
Jeto's architecture was built around a different mental model: treating saved content as ingredients for personalised media, not static archives. The system would understand context, relevance, and user preferences to create consumable experiences rather than storage systems.
Contextual social sharing addressed another broken pattern: the binary choice between public broadcast and private hoarding. Jeto enabled granular visibility controls, sharing specific collections with relevant communities while maintaining private experimentation spaces. Think Instagram's Close Friends feature, but for links of curiosity.
We also built ambient intelligence to power automatic organisation without rigid collections. Rather than forcing users into predetermined folder structures, the system would surface connections between saved items, suggest relevant collections, and predict optimal consumption timing based on historical patterns.
Our flagship feature was Smart Audio Stacks, AI-curated audio streams that transformed reading lists into commute-friendly audio snippets. The technical challenge wasn't just text-to-speech conversion but understanding narrative flow, topic curation, and optimal episode length based on user personalisation.
The vision resonated immediately with a small early group of friends and enthusiasts. We were even approached for a university-specific implementation that would let students save links to their subjects and articles and recall them later. This gave us the validation we needed to push ahead. We'd identified a genuine gap that affected users in their daily workflows. The challenge would be translating this insight into sustainable technology.
Technical architecture: lessons in AI-first design
Building Jeto made us face a basic challenge in modern product development: the gap between what AI can do and what's actually affordable.
The stack that almost worked
Next.js with React formed our web foundation. This turned out to be a good call since AI features required serious server-side processing, and the framework's edge computing capabilities helped us minimise latency in our content processing pipeline.
Supabase deserves a mention as our backend choice. For designers transitioning into full-stack development, Supabase's visual database interface removes traditional barriers to backend complexity. Row-level security policies felt intuitive rather than intimidating. Real-time subscriptions enabled social features without complex WebSocket implementations.
The decision to use Drizzle ORM with TypeScript eliminated entire categories of runtime errors. Auto-generated types from database schemas meant UI components could never display undefined data, which mattered a lot when AI processing introduces uncertainty into content structures.
The AI processing reality
Our LangChain implementation revealed the hidden complexity of production AI workflows. Academic tutorials make LLM integration appear straightforward, but production requires handling rate limits, context window optimisation, and graceful degradation when models are unavailable.
Content processing pipelines proved more nuanced than anticipated. Mozilla's Readability API works brilliantly for clean articles but struggles with modern web complexities: JavaScript-heavy sites, dynamic content loading, and paywall detection. We built fallback systems using headless browser automation, which dramatically increased infrastructure costs.
Vector databases for semantic search were another learning curve. Understanding embedding models, similarity thresholds, and retrieval strategies required digging into machine learning concepts that aren't typically part of product design education.
The architecture started to come together nicely in our staging environment. Eventually, as we started testing with several hundred users, we saw a different story entirely. More on that later.
Design philosophy: simplicity through intelligence
From a design perspective, Jeto was everything I'd learned about creating intuitive user experiences, put to the test. The challenge was making powerful AI features feel effortless rather than overwhelming.
The onboarding experience
We designed a progressive onboarding flow that demonstrated value before asking for commitment. New users could start saving links immediately, with smart suggestions appearing based on their initial selections. The AI would begin learning preferences from the first interaction, making each subsequent recommendation more relevant.
The interface prioritised visual browsing over text lists. Each saved link became a rich card showing the article's key image, auto-generated tags, estimated reading time, and AI-generated summary snippet. Users could preview content without clicking through, making it easier to rediscover forgotten saves.
Audio-first interactions
The audio features required entirely new interaction patterns. We designed a simple queue system where users could add articles to their "stack" throughout the day, then convert the entire collection into a playlist for their commute home.
Playback controls needed to work seamlessly across web and mobile, with features like speed adjustment, chapter markers for different articles, and the ability to save audio snippets for later sharing. We even planned integration with CarPlay and Android Auto for hands-free consumption.
Social features with intent
Unlike platforms that default to public sharing, Jeto's social features were designed around intentional curation. Users would create themed collections, "Weekend Reading," "Design Inspiration," "Tech Industry Analysis", and choose which ones to share publicly.
The "forking" feature allowed users to clone someone else's public collection as a starting point for their own curation. This created a middle ground between passive consumption and active creation, encouraging users to engage with content more thoughtfully.
What we actually built
Despite never shipping publicly, we built substantial functionality across web and mobile platforms. The scope was ambitious, perhaps too ambitious for a bootstrapped team.
Core bookmarking engine
The foundation worked beautifully. Users could save links from anywhere using browser extensions or by sharing to the mobile app. Our content parsing pipeline extracted article text, images, and metadata with impressive accuracy. The AI summarisation produced consistently useful previews that helped users decide whether to read or listen to full articles.
Intelligent tagging exceeded our expectations. The system could identify topics, sentiment, reading difficulty, and content type with remarkable precision. Users reported feeling like the app "understood" their interests in ways that manual organisation never achieved.
Mobile experience
Sasi's iOS development created an app that felt genuinely native. SwiftUI enabled smooth animations and responsive interactions that made browsing saved content feel effortless. Offline functionality worked seamlessly, users could queue articles while on WiFi and listen during commutes without connectivity concerns.
The mobile reading experience included features like adjustable typography, dark mode support, and gesture-based navigation. We paid particular attention to one-handed use, ensuring core functionality remained accessible even during crowded commutes.
Social and sharing features
Our collection system allowed sophisticated content organisation without overwhelming complexity. Users could create public showcases of their interests while keeping personal saves private. The interface made privacy settings clear and easy to modify.
The sharing features generated beautifully formatted cards suitable for social media platforms. Users could share individual articles, entire collections, or audio snippets with consistent visual branding that drove traffic back to Jeto.
Audio generation pipeline
Our audio features were the most technically challenging part of the platform. We built a complete pipeline from text processing to final audio delivery, with impressive results in our testing environment.
The AI-generated summaries struck the right balance between brevity and completeness. A typical 2,000-word article became a 3-4 minute audio summary that preserved key insights while remaining engaging for audio consumption. Users could adjust summary length and detail level based on their available time.
Voice quality was a particular focus. We experimented with multiple text-to-speech engines before settling on solutions that produced natural, engaging narration. The difference between robotic TTS and human-like speech mattered a lot for extended listening sessions.
The cost curve that broke our business model
The moment our financial projections went from optimistic to catastrophic is worth examining closely. It shows why so many promising AI applications struggle to achieve sustainable unit economics.
The optimism trap
Initial forecasting suffered from what I now recognise as AI economics denial, the assumption that rapidly improving technology would automatically translate to rapidly decreasing costs. We budgeted A$0 for AI hosting, expecting free tiers and open-source solutions to handle our processing needs indefinitely.
Reality arrived in the form of A$0.006 per second for quality audio generation. That number changed everything about our business model. A ten-minute audio summary, our core value proposition, cost A$3.60 to generate. Scale these numbers to meaningful user engagement, and the economics become unsustainable quickly.
The progression tells the story:
- Single user, moderate usage: A$45,372 annually
- 1,000 users at similar engagement: A$4.5 million annually
- Break-even scenarios: Required either massive scale or premium pricing that limited addressable market
Beyond the obvious costs
AI processing is only the visible portion of true infrastructure costs. Supporting real-time social features, content synchronisation, and audio delivery required more on top:
Content processing ran A$1,600 annually for web scraping services that could handle JavaScript-heavy sites and paywall detection. Database hosting was A$1,000 for Supabase Pro to support user growth projections. Content delivery cost A$400 for Vercel Pro to handle global audio distribution. Development infrastructure added another A$600 for design tools and A$110 for cloud platform experimentation.
These supporting costs, individually reasonable, compound into significant operational overhead before acquiring a single paying customer.
The unit economics reality
Consumer bookmarking applications face a basic problem: users consistently show low willingness to pay for content organisation tools. Successful platforms like Instapaper and Readwise charge A$60-120 annually, targeting niche professional audiences rather than general consumers.
Our freemium model assumptions collapsed under basic mathematics. Even optimistic conversion rates of 5-10% couldn't support the infrastructure costs required for AI-powered features. The disparity between user expectations (free with optional premium) and technical realities (expensive AI processing) created an unbridgeable gap.
This mismatch isn't unique to bookmarking. It's a broader problem facing AI-powered consumer applications. The technology enables compelling user experiences but at costs that often exceed consumer willingness to pay.
Why we stopped: time and money
The decision to shut down Jeto wasn't sudden, it emerged gradually as evidence mounted that our current approach wouldn't achieve the sustainability we needed.
The time reality
Building a comprehensive platform across web, mobile, and AI processing proved enormously time-intensive. As the sole designer and product lead, I found myself context-switching between user experience design, technical architecture decisions, business model planning, and basic project management.
What started as evening and weekend work began consuming entire days. The complexity of coordinating AI processing, real-time synchronisation, and cross-platform development stretched our small team beyond sustainable limits. Quality began suffering as we rushed to address technical debt and feature gaps.
The financial crossroads
Our commitment to bootstrapping Jeto ourselves meant every dollar spent came directly from personal funds. The AI cost realities forced an uncomfortable choice: compromise on the quality of our core differentiating feature or seek external investment much earlier than planned.
Neither option aligned with our values. Reducing audio quality to cheaper text-to-speech would undermine Jeto's central value proposition. Seeking investment would pressure us toward rapid scaling and monetisation rather than finding genuine product-market fit.
The financial projections painted a clear picture: even successful execution would require substantial runway to achieve profitability. Without external funding, we risked building something technically impressive but economically unsustainable.
Market reality
Research into the bookmarking and read-later market revealed additional challenges. Pocket, Instapaper, Matter, Readwise, had pivoted toward premium-first strategies with higher price points than we'd originally planned. This suggested our freemium approach might be fundamentally flawed for this market.
Users consistently showed low willingness to pay for bookmarking services, preferring free alternatives despite their limitations. The addressable market for premium bookmarking tools appeared smaller than our projections assumed.
Lessons learned: building for reality
Despite not shipping, Jeto taught us a lot about building AI-powered consumer applications and the realities of indie product development.
Technical insights
AI integration requires careful cost modelling from day one. The speed of AI capability improvement can make financial projections obsolete quickly, but costs remain a harsh constraint for consumer applications. Free tiers and optimistic forecasts are dangerous foundations for business planning.
Complexity compounds rapidly when building across multiple platforms with real-time synchronisation requirements. Each additional platform and feature multiplies testing, maintenance overhead, and potential failure points.
User experience quality can't be compromised for cost savings in core features. Cheap text-to-speech would have undermined our entire value proposition, but quality alternatives turned out to be unsustainably expensive at consumer price points.
Business learning
Bootstrap approaches require ruthless prioritisation. Our scope was too ambitious for available resources, leading to technical debt and feature gaps that compromised the overall experience.
Market research needs deeper financial analysis. Understanding user willingness to pay requires going beyond surveys to examine actual purchasing behaviour in adjacent markets.
Revenue model decisions shape product design more than you'd expect. Our freemium assumptions influenced every design decision, creating technical complexity that premium-first approaches might have avoided.
Design reflections
AI features need human-centred design that makes powerful capabilities feel approachable rather than overwhelming. The challenge isn't just building AI functionality but creating interfaces that help users understand and control it.
Audio interfaces require entirely new interaction patterns that don't translate directly from visual design principles. Designing for ears instead of eyes demanded different approaches to information hierarchy and user feedback.
Social features benefit from privacy-first design that defaults to personal use while enabling selective sharing. Users want control over their digital presence rather than platform-driven amplification.
Design principles that survived the pivot
Stepping away from Jeto gave us insights about AI-powered product design that go well beyond bookmarking applications.
The ambient intelligence question
AI features require different interaction paradigms than traditional software functionality. Users expect AI to work invisibly until they need conscious control. Jeto's tagging system succeeded when it felt like gentle suggestions; it failed when it demanded explicit decision-making.
This observation points to a simple design principle: AI should amplify human intent rather than replace human decision-making. The most successful features operated at the periphery of user attention, surfacing relevant information without demanding immediate action.
Audio interface design lessons
Building for audio consumption forced us to reconsider information hierarchy entirely. Visual design principles don't translate directly to audio experiences. Concepts like progressive disclosure and scannable content become meaningless when information flows linearly through time.
We discovered that audio interfaces require completely different approaches to navigation, content structure, and user feedback. The challenge isn't just converting visual designs to audio equivalents but reimagining how users interact with information when vision isn't available.
The privacy-first social model
Our approach to social features, private by default with intentional sharing, proved more resonant with users than anticipated. This suggests a broader trend toward contextual privacy controls that enable social discovery without sacrificing personal agency.
The "forking" concept for collections emerged as particularly powerful, enabling users to build upon others' curation without requiring ongoing social maintenance. This pattern could apply to many domain-specific applications beyond bookmarking.
Broader implications for AI product development
Jeto's development cycle has lessons for any team considering AI-powered consumer products.
Economic modelling must come before feature specification. AI capabilities improve rapidly, but costs remain stubbornly high for quality output. Business models built on assumptions of linear cost reduction often fail when confronted with actual usage patterns.
The gap between what AI can theoretically accomplish and what can be delivered sustainably at consumer price points is still wide. Managing this expectation gap requires careful feature prioritisation and transparent communication.
Infrastructure complexity also compounds quickly when combining AI processing with real-time social features. Each additional capability multiplies operational overhead and potential failure points.
What's next
The lessons from Jeto continue to inform our approach to product development at Happy Machines Company. We still want to build technology that genuinely improves human experience, just with more realistic expectations about timeline and scope.
For designers venturing into AI-powered products, I'd recommend starting with clear unit economics models that account for processing costs at meaningful scale. The technology is impressive and improving rapidly, but sustainable business models require honest assessment of current cost realities.
The bookmarking space still has opportunity, but success likely requires either substantially larger resources or much more focused approaches than we initially planned. Sometimes the most valuable product decision is recognising when to pivot toward problems better suited to what you actually have.
Building Jeto taught us that great products require more than technical capability and user-centred design, they demand business models that align technological possibilities with market realities. Sometimes the most educational projects are the ones you choose not to ship.